llamaperf

M5 Pro 24GB

APPLE · 24GB unified memory · 6 reports

Is this your card?

On the M5 Pro 24GB, DeepSeek-Coder-V2-Lite 16B · 2.4B active at 4-bit, the model most run on it, generates a median of 125.8 tokens per second from one plain run (community reports, as of 22 Nov 2026).

As of 22 Nov 2026, the models most run on the M5 Pro 24GB, each one size at one quant level, with the median generation speed of its plain runs (one device, one request, no speculative decoding):

Bars are medians on one scale, 0 to 125.8 tokens/s. Press ▶ to watch text arrive at a row's speed.

Engines people use on it: Inco Splash 1 (reports) · MLX 1 (reports) · llama.cpp 1 (reports)

Run models on your M5 Pro 24GB? Add your numbers.

Your coding agent can run a speed test on your machine and send it here with a free account. Or paste a result you've already posted.

Reported performance on the M5 Pro 24GB

These reports mix GPU counts and memory setups, so a model listed here may have needed extra cards or CPU offloading. The calculator estimates which models fit in 24 GB of unified memory. Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.

Speeds are as each source reported them. GPU count, offloading and concurrent requests change them, and prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing two reports.

Tone: positiveApple hardware
generation:
61.8 tokens/s
quant:
4bit (MLX)
visionlong-contextagentic

User reports Qwen3.8-27B at 61.8 tok/s on an M5 Pro 24GB Mac, up from 57.3 tok/s with draft-tree speculative decoding off. Setup is the native Rust/Metal JustFit 2.6 engine with stock MLX 4-bit weights and Inco AI's DFlash2 draft, on a 24 GB Apple Silicon machine. The 7-prompt greedy suite also gave 40.9 tok/s on M6 (35.6 chain) and 50.9-51.3 tok/s on M4 Pro with a fixed 8-row tree. On NVIDIA SPEED-Bench at 16K prompts the M5 Pro median was 59.5 tok/s, M6 38.5 tok/s and M4 Pro 48.3 tok/s. Dual 128K+32K requests peaked at 21,563 MiB on M5 Pro.

Nov 22, 2026
Tone: positiveApple hardware
generation:
14.7 tokens/s
prompt processing (prefill):
374.1 tokens/s
quant:
IQ3_XXS (GGUF)
agentic

User benchmarks Qwen3.8-Flash-Next 125B at 14.69 tok/s decode on an M5 Pro 24GB, with 374.15 tok/s prefill on a 6,060-token prompt. Setup is the MoEShift backend with an IQ3_XXS quant (76GB on disk) and SSD offloading, peak process memory 8.5GB. Medians of 3 runs with prompt caching disabled. On the same setup an M4 Pro 24GB reached 9.84 tok/s short-prompt decode, 199.09 tok/s prefill and 9.57 tok/s long-prompt decode; long-prompt inference took 55.1s on the M5 Pro versus 83.9s on the M4 Pro.

Nov 9, 2026

Qwen3.8 27B

M5 Pro 24GB · Inco Splash · 120,000 ctx

Tone: positiveApple hardware
generation:
55.0 tokens/s
prompt processing (prefill):
448.0 tokens/s
quant:
UD-IQ3_XXS (GGUF)
agentictool-use

User reports Qwen3.8-27B at about 55 t/s generation and 448 t/s prefill on an M5 Pro MacBook Pro with 24GB unified memory. Setup is the Inco Splash engine (v1.4.1) with the UD-IQ3_XXS GGUF (10.9GB), capped at 18GB memory for around 120k context. User's own KLD measurements give 89.5% top-1 agreement with mean KLD 0.0542, and they say Safari and Discord still run comfortably alongside it.

Oct 27, 2026
Tone: positiveApple hardware
generation:
73.9 tokens/s
quant:
IQ3_XXS (GGUF)

User reports decode speed improvements on Qwen3-30B-A3B with IQ3_XXS quantization on an M5 MacBook Pro 24GB. The decode speed increased from 65.6 to 73.9 tok/s using a llama.cpp PR for Metal optimization. Setup uses llama.cpp with IQ3_XXS quantization. The user also tested on an M1 Pro 32GB. The user is looking for testers for Qwen3.8-Flash-Next and provides benchmark commands.

Oct 6, 2026
Tone: positiveApple hardware
generation:
125.8 tokens/s
prompt processing (prefill):
363.9 tokens/s
quant:
4bit (MLX)
codingagentic

User benchmarks DeepSeek-Coder-V2-Lite-Instruct at 125.8 tok/s generation and 363.9 tok/s prompt processing on an Apple M5 Pro with 24GB unified memory. Setup is MLX with a 4bit quant (8.2GB) on a single machine, running a fixed coding task capped at 1500 output tokens. The model produced correct working code; the user notes that 24GB has a hard ceiling around 20GB models and that background downloads measurably affect tok/s.

Oct 3, 2026
Tone: positiveApple hardware
generation:
4.8 tokens/s
quant:
2-bit dynamic

User reports DeepSeek-V4-Flash 284B-A13B at 4.8 tok/s on a 24 GB M5 Pro with about 5.3 GB of memory in use. Setup is Mference, an open-source engine. The user also runs Gemma 4 26B-A4B and Qwen 3.6 35B-A3B.

Sep 7, 2026

Get a weekly email of new M5 Pro 24GB reports and newly released models that fit it.

Email me new reports

Reported configurations

Individual observations from this page, not expected speeds or a ranking. Select a model to open its report, with the source and offloading conditions. Reported context may be a limit rather than actual input length.

Model / hardwareQuant / engineReported contextGeneration
Qwen3.8 27B
M5 Pro 24GB
4bit
JustFit
Not reported61.8 tokens/s
Qwen3.8 125B (6B active) Flash-Next
M5 Pro 24GB
IQ3_XXS
MoEShift
6,06014.7 tokens/s
Qwen3.8 27B
M5 Pro 24GB
UD-IQ3_XXS
Inco Splash
120,00055.0 tokens/s
Qwen3 30B (3B active)
M5 Pro 24GB
IQ3_XXS
llama.cpp
Not reported73.9 tokens/s
DeepSeek-Coder-V2-Lite 16B (2.4B active)
M5 Pro 24GB
4bit
MLX
Not reported125.8 tokens/s
DeepSeek V4 Flash 284B (13B active)
M5 Pro 24GB
2-bit dynamic
Mference
Not reported4.8 tokens/s

Frequently asked

How fast is the M5 Pro 24GB for local LLMs?

On the M5 Pro 24GB, DeepSeek-Coder-V2-Lite 16B · 2.4B active at 4-bit, the model most run on it, generates a median of 125.8 tokens per second from one plain run (community reports, as of 22 Nov 2026). Also reported: Qwen3 30B · 3B active at 3-bit (73.9 t/s, one plain run).

What is the biggest model reported running on the M5 Pro 24GB?

The largest model with a plain generation run on the M5 Pro 24GB is Qwen3 30B · 3B active at 3-bit (73.9 t/s, one plain run), as of 22 Nov 2026. Plain means one device, one request, no speculative decoding, so a run split across several cards does not count.